Wetlands ( IF 1.8 ) Pub Date : 2022-04-27 , DOI: 10.1007/s13157-022-01558-2
Mohammed T. Zaki 1 , Omar I. Abdul-Aziz 1
|
Prediction of wetland greenhouse gas (GHG) fluxes has been a challenging undertaking. Machine learning techniques such as the artificial neural network (ANN) has a strong potential to provide high quality predictions of the wetland GHG fluxes. We developed eight different ANN models and investigated their suitability to predict the major GHG fluxes (CO2 and CH4) in coastal salt marshes (dominated by Spartina alterniflora) of Waquoit Bay, Massachusetts, USA. Based on the dominant environmental drivers, the daytime net uptake fluxes of CO2 were predicted as a function of photosynthetically active radiation, soil temperature (ST), and porewater salinity (SS). The net emission fluxes of CH4 were predicted as a function of ST and SS. Our models with the radial basis function neural network (RBNN) provided the most accurate and least-biased predictions of the net CO2 uptake (Nash-Sutcliffe Efficiency, NSE = 0.98) and CH4 emission (NSE = 0.90-0.92). The linear layer neural network generated the least successful and most biased predictions of the GHG fluxes (NSE = 0.48-0.80). Other ANNs, including the commonly-used feed forward neural network (FFNN), provided less accurate and more biased predictions of the CO2 (NSE = 0.86-0.97) and CH4 (NSE = 0.73-0.89) fluxes than the RBNN. We, therefore, recommend using RBNN as the first choice and FFNN (or its variant) as the second choice for predicting the GHG fluxes in coastal salt marshes. Our findings and tools would help derive plausible scenarios and guidelines for restoration, monitoring, and maintenance of coastal salt marshes in the U.S. and beyond.
中文翻译:
![](https://scdn.x-mol.com/jcss/images/paperTranslation.png)
使用人工神经网络预测沿海盐沼中的温室气体通量
湿地温室气体 (GHG) 通量的预测一直是一项具有挑战性的工作。人工神经网络 (ANN) 等机器学习技术具有提供湿地温室气体通量的高质量预测的强大潜力。我们开发了八种不同的人工神经网络模型,并研究了它们对预测美国马萨诸塞州瓦奎特湾沿海盐沼(以互花米草为主)中主要温室气体通量(CO 2和 CH 4 )的适用性。基于主要的环境驱动因素,CO 2的白天净吸收通量被预测为光合有效辐射、土壤温度 (ST) 和孔隙水盐度 (SS) 的函数。CH 4的净排放通量被预测为 ST 和 SS 的函数。我们的径向基函数神经网络 (RBNN) 模型对 CO 2净吸收(Nash-Sutcliffe 效率,NSE = 0.98)和 CH 4排放(NSE = 0.90-0.92)提供了最准确和最小偏差的预测。线性层神经网络生成的温室气体通量预测最不成功且偏差最大(NSE = 0.48-0.80)。其他人工神经网络,包括常用的前馈神经网络 (FFNN),对 CO 2 (NSE = 0.86-0.97) 和 CH 4的预测精度较低且偏差较大(NSE = 0.73-0.89) 通量高于 RBNN。因此,我们推荐使用 RBNN 作为预测沿海盐沼温室气体通量的首选,FFNN(或其变体)作为第二选择。我们的发现和工具将有助于为美国及其他地区的沿海盐沼的恢复、监测和维护得出合理的情景和指导方针。